Improving Pulse Rate Measurements during Random Motion ... - MDPI

sensors Article

Improving Pulse Rate Measurements during Random Motion Using a Wearable Multichannel Reflectance Photoplethysmograph Kristen M. Warren 1 , Joshua R. Harvey 1 , Ki H. Chon 2 and Yitzhak Mendelson 1, * 1 2

*

Department of Biomedical Engineering, Worcester Polytechnic Institute, Worcester, MA 01605, USA; [email protected] (K.M.W.); [email protected] (J.R.H.) Department of Biomedical Engineering, University of Connecticut, Storrs, CT 06269, USA; [email protected] Correspondence: [email protected]; Tel.: +1-508-831-5103

Academic Editor: Oliver Amft Received: 2 January 2016; Accepted: 1 March 2016; Published: 7 March 2016

Abstract: Photoplethysmographic (PPG) waveforms are used to acquire pulse rate (PR) measurements from pulsatile arterial blood volume. PPG waveforms are highly susceptible to motion artifacts (MA), limiting the implementation of PR measurements in mobile physiological monitoring devices. Previous studies have shown that multichannel photoplethysmograms can successfully acquire diverse signal information during simple, repetitive motion, leading to differences in motion tolerance across channels. In this paper, we investigate the performance of a custom-built multichannel forehead-mounted photoplethysmographic sensor under a variety of intense motion artifacts. We introduce an advanced multichannel template-matching algorithm that chooses the channel with the least motion artifact to calculate PR for each time instant. We show that for a wide variety of random motion, channels respond differently to motion artifacts, and the multichannel estimate outperforms single-channel estimates in terms of motion tolerance, signal quality, and PR errors. We have acquired 31 data sets consisting of PPG waveforms corrupted by random motion and show that the accuracy of PR measurements achieved was increased by up to 2.7 bpm when the multichannel-switching algorithm was compared to individual channels. The percentage of PR measurements with error ď 5 bpm during motion increased by 18.9% when the multichannel switching algorithm was compared to the mean PR from all channels. Moreover, our algorithm enables automatic selection of the best signal fidelity channel at each time point among the multichannel PPG data. Keywords: motion artifacts; multichannel photoplethysmograph; multichannel template matching; pulse rate; wearable sensor; pulse oximeter

1. Introduction Pulse oximetry uses light absorption to measure arterial blood oxygen saturation (SpO2 ) and pulse rate (PR) from photoplethysmographic (PPG) signals. Pulse oximeters detect a pulsatile signal that is normally only a small percentage of the total PPG signal. Therefore, any transient motion of the sensor relative to the skin, such as during exercise, can cause a significant artifact in the optical measurement. Furthermore, if these artifacts mimic a heartbeat, the instrument may not be able to differentiate between the pulsations that are due to motion artifacts (MA) and normal arterial pulsations, thereby distorting the PPG waveforms and causing false or erroneous PR readings. The primary cause of MA in pulse oximetry is predominantly due to changes in the light path during sensor movements [1]. Pulse oximetry is widely used in hospitals where motion artifacts are generally less pronounced compared to mobile health applications. For instance, during patient transport, bouncing of the vehicle

Sensors 2016, 16, 342; doi:10.3390/s16030342

www.mdpi.com/journal/sensors

Sensors 2016, 16, 342

2 of 18

can cause the probe to be displaced and temporarily lose the PPG signal. Thus, the stability of the PPG signal is essential to facilitate reliable vital sign monitoring. Motion artifacts are difficult to filter out since they do not have a predetermined frequency range and their spectral content often overlaps with the frequency band of the PPG waveform. If motion artifact persists long enough and has a frequency in the range of normal PR, the calculated PR can be highly inaccurate. Clinicians have cited motion artifacts in pulse oximetry as the most common cause of false alarms, loss of signal, and inaccurate readings [2]. The primary approach to reduce the effects of motion artifact is the implementation of software-based algorithms that attempt to extract a clean PPG waveform from the motion-corrupted PPG signal [3–18]. Conventional signal filtering is helpful, but not very effective when MA has no predetermined frequency band. Particularly, when the noise is near or shares the same frequency band as the signal components of interest, this technique can suppress the desired signals. Baseline subtraction, use of frequency banks, moving average filtering, and removal of corrupted signal segments have shown improvement against MA in some cases, but are not robust against motion which may have varied dynamics [19–22]. Numerous studies have investigated the use of adaptive noise cancellation (ANC) as an alternative approach to selectively filter out MA based on a specified reference signal. Reference signals used include: on-board accelerometers [3–6], a reference signal synthesized from the motion corrupted PPG signal [7–10], and a reference signal measured by an adjacent photoelectric device [11,12]. However, a separate reference signal is not always an accurate representation of the signal corruption, which can lead to unintentional filtering of a portion of the relevant PPG waveform. Alternate algorithms have been developed to extract the clean PPG waveform from the motion-corrupted signal based on fundamental components of the PPG signal. These methods include: principle component analysis (PCA) [13], independent component analysis (ICA) [14–16], and singular spectral analysis (SSA) [17]. Most recently, algorithms based on filtering out the motion frequency as taken from the accelerometer spectra have been useful in separating motion signal from PPG signal [23,24]. These algorithms have been proven somewhat effective during motion, but they are not currently optimized for multiple channel recordings and photoplethysmogram sensors only consist of a single pair of red (RD) and infrared (IR) light emitting diodes (LEDs) and a single photodetector (PD). When the single channel is too corrupted to reconstruct, or when the motion frequency overlaps with the PR frequency preventing motion from being filtered out, PR and SpO2 information may be lost, leading to dropouts during monitoring. Multichannel devices have been used to combat motion artifacts by capturing multiple PPG waveforms simultaneously. An 8-channel PPG sensor placed on the sternum was developed that uses PCA in the frequency domain to find the most likely SpO2 estimation [25]. These investigators found that multichannel SpO2 estimates were more robust than single channel SpO2 estimates. A 3-channel reflectance in-ear sensor was developed and tested during standing, sitting, and walking [26]. An adaptive notch filter was implemented at the motion frequency to reduce noise contribution. These investigators found that motion-induced current was channel-specific, and that the channel with the highest power around the PR frequency varied between experimental runs. These studies showed that multichannel pulse oximetry is advantageous over single channel pulse oximetry in obtaining diverse signal information during low-motion artifact conditions. Studies have attempted to better characterize the effects of motion artifact in pulse oximetry, and have shown that although all types of motion lead to measurement errors, a majority of errors are generated by intense, aperiodic, random movements [27]. Previously, we have shown that in a six-channel prototype reflectance-based forehead pulse oximeter, during short up-down, left-right, and circular head motion, channels responded differently to motion [28]. However, in the aforementioned work, we did not investigate whether the added complexity of a multichannel PPG approach actually produces measurable benefits compared to a single channel sensor and we also did not address the question of how to fuse the data captured by multiple channels to obtain better results than when analyzing data captured by a single channel PPG sensor. While it might seem intuitive that multiple

Sensors 2016, 16, 16, 342 342 Sensors 2016,

of 18 18 33 of

intuitive that multiple channels are superior to a single channel, the main challenge lies in finding suitable to to actually potential. Thus, how multichannel channels methods are superior a singleleverage channel, this the main challenge lies examining in finding suitable methods to photoplethysmography responds a wide how variety of motionphotoplethysmography in comparison to the more actually leverage this potential. Thus,toexamining multichannel responds conventional single-channel approach would help to further assess the benefit of this unique would sensor to a wide variety of motion in comparison to the more conventional single-channel approach design and assist inthe developing advanced that make use of help to further assess benefit of this uniquesignal sensor extraction design andalgorithms assist in developing advanced multichannel waveforms tothat improve motion tolerance inwaveforms photoplethysmography including pulse signal extraction algorithms make use of multichannel to improve motion tolerance in oximetry. photoplethysmography including pulse oximetry. In this paper, we explore the benefit benefit of of using using aa customized customized forehead-mounted forehead-mounted multichannel photoplethysmographic sensorcomprising comprisingsixsix photodetectors ofand redinfrared and infrared photoplethysmographic sensor photodetectors andand twotwo pairspairs of red LEDs, LEDs, and investigate the performance of wearable this newsensor wearable sensor under a variety of random and investigate the performance of this new under a variety of random motion. Since motion. Since current are not necessarily able tomeasurement select which measurement of PR isbased most current algorithms arealgorithms not necessarily able to select which of PR is most correct correct based on signal weadvanced introducemultichannel-switching an advanced multichannel-switching algorithm on signal corruption, wecorruption, introduce an algorithm that selects the that selects thethe channel with theofleast amount of motion artifact calculate PRshow everythat 2 s.for Wea show channel with least amount motion artifact to calculate PR to every 2 s. We wide that forofarandom wide variety of random motion, channels respondand differently to motion, andfrom the variety motion, channels respond differently to motion, the multichannel estimate multichannel estimate from the recorded sensor array outperforms estimates in signal terms the recorded sensor array outperforms single-channel estimates in single-channel terms of motion tolerance, of motion tolerance, signal quality, and PR errors. quality, and PR errors. 2. Experimental ExperimentalSection Section 2.

2.1. Device Description and Experimental Setup 2.1.1. Sensor Description 2.1.1. Sensor Description We have developed a customized forehead-mounted, wearable multichannel We have developed a customized forehead-mounted, wearable multichannel photoplethysmographic photoplethysmographic (MCP) sensor operating in reflectance mode, as shown in Figure 1. Six (MCP) sensor operating in reflectance mode, as shown in Figure 1. Six surface-mounted Si photodetectors surface-mounted Si photodetectors (PD), each having an active area of 2.65 mm2, are positioned (PD), each having an active area of 2.65 mm2 , are positioned concentrically as a symmetrical array around concentrically as a symmetrical array around two pairs of red (660 nm) and IR (940 nm) light two pairs of red (660 nm) and IR (940 nm) light emitting diodes (LEDs) at an equidistant separation emitting diodes (LEDs) at an equidistant separation distance of 10 mm [29]. Both pairs of Red and IR distance of 10 mm [29]. Both pairs of Red and IR LEDs were turned on in pairs, at the same time, LEDs were turned on in pairs, at the same time, alternating between Red LEDs and IR LEDs. Two alternating between Red LEDs and IR LEDs. Two LEDs of each wavelength were used to increase total LEDs of each wavelength were used to increase total brightness. An opaque ring was incorporated brightness. An opaque ring was incorporated to minimize direct light shunting between the LEDs and the to minimize direct light shunting between the LEDs and the adjacent PD array. adjacent PD array.

Figure 1. Six-photodetector (6PD) Mounted Forehead Mounted Multichannel Reflectance-Mode Multichannel Figure 1. Six-photodetector (6PD) Forehead Reflectance-Mode Photoplethysmographic (MCP) Sensor. Photoplethysmographic (MCP) Sensor.

The asas a The sensor sensor and and battery battery are areenclosed enclosedin inaaplastic plasticcasing casingand andattached attachedtotoan anelastic elasticband bandworn worn headband, allowing the sensor to rest comfortably on the forehead. The sensor is also equipped with a headband, allowing the sensor to rest comfortably on the forehead. The sensor is also equipped an analog tri-axial MEMS accelerometer (Acc) to to detect movement with an analog tri-axial MEMS accelerometer (Acc) detect movementwith withrespect respectto to gravitational gravitational acceleration andtotoprovide provide a reliable movement reference. the sensor is on placed on the acceleration and a reliable movement reference. WhenWhen the sensor is placed the forehead, forehead, the x-direction of the Acc corresponds to motion perpendicular to the transverse plane, the the x-direction of the Acc corresponds to motion perpendicular to the transverse plane, the y-direction y-direction perpendicular to sagittal plane, and theperpendicular z-direction perpendicular the coronal plane. perpendicular to sagittal plane, and the z-direction to the coronalto plane. Previous data Previous data have shown that variations in sensor position and vasculature heterogeneity the have shown that variations in sensor position and vasculature heterogeneity of the underlying of tissue underlying tissue can cause measurement errors, as well as light diffusion by the subcutaneous can cause measurement errors, as well as light diffusion by the subcutaneous tissues predominantly tissues predominantly in the to direction perpendicular to the emitting surfacemotions of the LEDs [28]. in the direction perpendicular the emitting surface of the LEDs [28]. Different and sensor Different motions and sensor contact pressure change the optical coupling between the sensor and

Sensors 2016, 16, 342 Sensors 2016, 16, 342

4 of 18 4 of 18

contact pressure change the optical coupling between sensor and underlying skin, yielding the underlying skin, yielding 6 independent PPG the channels withthe different motion-corrupted 6waveform independent PPG channels with different motion-corrupted waveform characteristics. characteristics. 2.1.2. 2.1.2. Data Data Collection Collection Data Worcester Data were were collected collected from from 15 15 healthy healthy volunteers volunteers between between the the ages ages of of 23 23 and and 30. 30. Worcester Polytechnic Institute IRB approved the study protocol and informed consent was required Polytechnic Institute IRB approved the study protocol and informed consent was required by by all all subjects prior to data recording. subjects prior to data recording. Subjects Subjects were were asked asked to to bounce bounce on on an an exercise exercise ball ball while while wearing wearing the the 6PD 6PD MCP MCP forehead forehead sensor sensor ® and CA, USA) and aa reference reference Masimo-57 Masimo-57 Radical Radical (Masimo (Masimo SET SET® ,, Masimo Masimo Corporation, Corporation, CA, USA) finger finger type type transmittance pulse oximeter that was kept motionless by resting the left hand on a table, as shown transmittance pulse oximeter that was kept motionless by resting the left hand on a table, as shown in 2. Each Eachsubject subjectwas wasasked askedtotoalternate alternate between 3 min 5 min of bouncing on in Figure Figure 2. between 3 min of of restrest andand 5 min of bouncing on the the exercise a total of min. 19 min. Subjects were instructed to move in any particular way, exercise ballball for for a total of 19 Subjects were notnot instructed to move in any particular way, or or to to move at any given frequency, yielding data sets with different types of movement artifacts. move at any given frequency, yielding data sets with different types of movement artifacts. This This protocol to introduce motion artifacts with a variety frequencies and amplitudes protocol was wasimplemented implemented to introduce motion artifacts with aofvariety of frequencies and and an increasing and decreasing PR. Six pairs of PPG waveforms corrupted by random motion amplitudes and an increasing and decreasing PR. Six pairs of PPG waveforms corrupted by random artifacts were obtained from the forehead-mounted MCP sensor. PPG waveforms from the MCP motion artifacts were obtained from the forehead-mounted MCP sensor. PPG waveforms from the were at 80 Hz, was sufficient to recover the shape of the signal. Reference PR MCP sampled were sampled at 80which Hz, which was sufficient to recover the shape ofPPG the PPG signal. Reference measurements were obtained from the Masimo pulse oximeter every 2 s. All data were captured PR measurements were obtained from the Masimo pulse oximeter every 2 s. All data were captured simultaneously by a a PC simultaneously by PC and and processed processed offline offline with with MATLAB. MATLAB. Data Data from from the the MCP MCP and and the the Masimo Masimo reference sensor were aligned by matching PR trends during rest for all data sets. reference sensor were aligned by matching PR trends during rest for all data sets.

Figure 2. Experimental setup for generating random motion. Figure 2. Experimental setup for generating random motion.

2.2. Methodology 2.2. Methodology The study was designed to test the response of the MCP to motion artifacts with a variety of The study was designed to test the response of the MCP to motion artifacts with a variety of frequencies and amplitudes. To quantify subjects’ motion, the RMS values from the on-board frequencies and amplitudes. To quantify subjects’ motion, the RMS values from the on-board tri-axial tri-axial accelerometer were calculated for each data set. Furthermore, because PR calculations are accelerometer were calculated for each data set. Furthermore, because PR calculations are highly highly dependent on the frequency content of the IR PPG waveform, the power spectral density dependent on the frequency content of the IR PPG waveform, the power spectral density (PSD) was (PSD) was calculated for each data set, and for each IR channel, to determine how motion calculated for each data set, and for each IR channel, to determine how motion frequencies affect frequencies affect the multichannel PPG waveforms. Oxygenated hemoglobin absorbs comparably the multichannel PPG waveforms. Oxygenated hemoglobin absorbs comparably less red light than less red light than IR, so the AC component of the red PPG signal is smaller than the IR signal and IR, so the AC component of the red PPG signal is smaller than the IR signal and the responsivity of the responsivity of the PD is higher in the IR region of the spectrum. Therefore, the IR PPG the PD is higher in the IR region of the spectrum. Therefore, the IR PPG waveforms are used for PR waveforms are used for PR measurements due to the generally better signal quality. In order to quantify instantaneous noise levels, a multichannel template-matching algorithm was developed

Sensors 2016, 16, 342

Sensors 2016, 16, 342

5 of 18

5 of 18

measurements due to the generally better signal quality. In order to quantify instantaneous noise that matches beats intemplate-matching a specified window to an average template PPG levels, a multichannel algorithm was developed thatrepresentative matches beats of in aclean specified morphology. window to an average template representative of clean PPG morphology.

2.2.1. Motion Motion Quantification Quantification 2.2.1. Motion across across channels channels was was compared compared for for all all data data sets sets using using the the following following parameters: parameters: Motion multichannel template-matching noise level (MCNL), RMS accelerometer amplitude, and PSD PSD multichannel template-matching noise level (MCNL), RMS accelerometer amplitude, and amplitude at the motion frequency. amplitude at the motion frequency. Multichannel Template TemplateMatching Matching Multichannel Toquantify quantifythethe noise in channel each channel at a particular point, wea multichannel developed a To noise levellevel in each at a particular time point,time we developed multichannel template-matching algorithm. Algorithms have been developed based on creating template-matching algorithm. Algorithms have been developed based on creating a template froma template from single channel IR PPG waveforms using an average of beats over a specified window single channel IR PPG waveforms using an average of beats over a specified window [30]. In our [30]. In oursix algorithm, six IR PPGare waveforms are used to create a template overperiod, a single time period, algorithm, IR PPG waveforms used to create a template over a single time allowing more allowing more robust template First, the raw PPG were with digitally filtered with robust template formation. First,formation. the raw PPG waveforms werewaveforms digitally filtered a zero-phase 6th a zero-phase 0.5 to 12 Hz Butterworth band-pass filter. The six filtered IR components order, 0.5 to 126th Hzorder, Butterworth band-pass filter. The six filtered IR components served as inputs to served as inputs template-matching to the multichannelalgorithm. template-matching algorithm. Initialization of this the multichannel Initialization of this algorithm assumes thatalgorithm the data assumes the rest, datawith startsclean out PPG during rest, with The clean PPG waveforms. is divided starts out that during waveforms. algorithm is dividedThe intoalgorithm two separate parts, into two formation separate parts, template formation and multichannel Both parts of the template and multichannel noise calculation. Both partsnoise of thecalculation. algorithm are depicted in the algorithmin are depicted flowchart Figure 3. in the flowchart in Figure 3.

Figure 3. 3. Processing the multichannel multichannel template-matching template-matching algorithm algorithm to to obtain obtain Figure Processing of the data with the Multichannel Noise Noise Level Level (MCNL) (MCNL) for for each each channel. channel. Multichannel

(a) Template Formation Peak–rough detection was performed on each of the six IR components over a 12 s window. The peak–trough detection was implemented by alternating between finding local minima and local maxima in a ±0.25 s window. The maximum or minimum found must be the absolute maximum or

Sensors 2016, 16, 342

6 of 18

(a) Template Formation Peak–rough detection was performed on each of the six IR components over a 12 s window. The peak–trough detection was implemented by alternating between finding local minima and local maxima in a ˘0.25 s window. The maximum or minimum found must be the absolute maximum or minimum in this 0.5 s window preventing features, such as the dicrotic notch, to be incorrectly calculated as a peak or trough. From this, an average peak-peak period was calculated across all channels and was defined by a variable L. Peak-peak period values that are outside of L ˘ 0.2 s were removed to include only clean beats in the beat period calculation. The average peak-peak period is updated by taking the average of the remaining peak-peak period values. Beats in each window across all channels were segmented at the indices of each trough. Beats across all channels for the entire 12 s of data were re-sampled to be the same length (L) using a cubic spline, and normalized to have a maximum amplitude of 1 and a minimum amplitude of 0. A temporary template is created from the average of all beats in the 12 s window. The correlation between this template and all beats in the current window was computed, and if the correlation of a beat in the current window with this template was less than 0.95, the beat was removed from further calculations. If the number of beats removed were more than 1/3 of the total number of beats in the current window, the template from the previous window was used. If less than 1/3 of the beats are removed, taking the average of the remaining “good” beats in the window forms a new template. The average correlation of all individual beats in the window is again calculated with the latest template. If the average correlation across all beats in the window with this template is less than 0.98, the template from the previous window is used. If the average correlation across all beats in the window is greater than 0.98, the template from the previous step is maintained. (b) Multichannel Noise Calculation With the final formation of the template for a given window, the template-matching algorithm systematically overlays beats from each individual channel, cuts off each beat at L/2 in order to capture only the systolic morphology of the PPG beats, which are more indicative of clean signals, and normalizes each beat such that the minimum amplitude is 0 and the maximum amplitude is 1. If the noise is such that there are no beats detected in a given channel, the multichannel noise level (MCNL) is set to 1. Otherwise, the correlation between the template and each beat in a given channel is calculated, and the overall MCNL is calculated as (1-C), where C is the average correlation between beats in a single channel and the template. This multichannel template-matching algorithm is implemented using a 2 s sliding window on the AC IR PPG data. Every 2 s, the channel with the lowest MCNL is selected to calculate PR. PR, in beats per minute (bpm), is calculated by dividing 60 by the average peak-to-peak period in seconds of the selected channel across the 12 s window. Accelerometer Amplitude The accelerometer signals measured during motion provide a way to quantify the amount of motion that was introduced into any particular data set. All accelerometer signals were filtered with the same filter we used to filter the PPG data, i.e., a 6th order zero-phase, 0.5 to 12 Hz Butterworth band-pass filter. The RMS value of all 3 axes of the on-board accelerometer was used to quantify motion, according to Equation (1). c AccelRMS “

1 ppAccelX q2 ` pAccelY q2 ` pAccelZ q2 q 3

(1)

Although these calculations do not translate directly into the level of noise introduced into the corrupted PPG waveforms during motion, they help to distinguish between high motion and low motion data sets based on the overall RMS values.

Sensors 2016, 16, 342

7 of 18

Motion Frequency Differences The power spectral density (PSD) of each PPG waveform represents the power of the PPG waveform at each frequency. The PSD of the AC IR PPG waveform in each channel is calculated using Welch’s periodogram. The PSD of each axis of the accelerometer was also calculated using the same method. The mean pulse rate (MPR) during motion was calculated according to the reference readings from the Masimo finger sensor during motion. The motion frequency was determined by using the accelerometer PSD, and then the amplitude of the PPG waveform PSD at the motion frequency was compared. 2.2.2. Pulse Rate Performance Metrics Pulse rate error was defined as the absolute error between the PR calculated by the multichannel device and the PR from the Masimo-57 reference sensor, according to Equation (3). ErrPR rbpms “ |PRMCP ´ PRMasimo |

(2)

This PR error was analyzed during motion using three parameters: accuracy, precision, and performance index. The Masimo-57 pulse oximeter claims a PR error tolerance of ˘5 bpm during motion [31]; thus, we used a ˘5 bpm tolerance for performance index to match the Masimo specifications during motion. Performance index was defined as the percentage of the measurements that have absolute relative PR errors lower than 5 bpm during motion. The percentage in performance index corresponds to the number of low-error measurements taken during motion. Accordingly, Performance Index r%s “

# measurements pErrPR ď 5 bpmq ˆ 100 total # of measurements

(3)

Accuracy was defined as the offset that a PR measurement has in relation to the reference device. In this paper, we consider accuracy as the mean absolute PR error during motion in relation to the measurements by the Masimo-57 reference sensor. Accordingly, Accuracy rbpms “ AVERAGE pErrPR q

(4)

Precision was defined as the ability to make consistent measurements, representative of the spread of measurements taken around the measured value. Hence, precision was calculated as the standard deviation of the absolute error taken in relation to the Masimo-57 reference sensor. Accordingly, Precision rbpms “ STD pErrPR q

(5)

To compare the multichannel estimates against the single channel estimates for each parameter, eight one-sided Student’s t-tests were performed: estimates from each individual channel were compared against the multichannel-switching estimate for each parameter. The mean and median PR for each data set were compared against the multichannel-switching estimate. A confidence value of 95% (α = 0.05) was used to find t-critical values for each t-test. 3. Results 3.1. Time-Domain PPG Waveform Differences During Motion The infrared PPG waveforms are similar during rest across all channels, and differ in levels of signal corruption between channels during motion. Figure 4 shows infrared PPG waveforms recorded during rest and motion from Data Set 10.

Sensors 2016, 16, 342

8 of 18

Sensors 2016, 16, 342 the pressure exerted by the headband to secure the sensor and the signal amplitude of the 8 of 18 However,

PPG waveform, amongst other variables, can affect the quality of the recorded PPG waveforms.

Sensors 2016, 16, 342

8 of 18

However, the pressure exerted by the headband to secure the sensor and the signal amplitude of the PPG waveform, amongst other variables, can affect the quality of the recorded PPG waveforms.

(a)

(b)

Figure 4. Infrared photoplethysmogram (PPG) waveform differences in all six channels recorded

Figure 4. Infrared photoplethysmogram (PPG) waveform differences in all six channels recorded during rest (a) and motion (b) in a 12-s window for Data Set 10. during rest (a) and motion (b) in a 12-s window for Data Set 10.

3.2. Accelerometer Range of Motion (b) an indirect measure of how The RMS amplitude of (a)the on-board tri-axial accelerometers provide much noise was introduced into the corrupted PPG waveforms in each data set, as summarized in Figure 4. Infrared photoplethysmogram (PPG) waveform differences in all six channels recorded Figure 5. during rest (a) and motion (b) in a 12-s window for Data Set 10.

Figure 5. Box and Whisker plot of root mean square (RMS) accelerometer amplitudes across all data sets during rest (left side of column) and motion (right side of column). RMS values were calculated using Equation (1). The edges of the box indicate the 25th and 75th percentiles, the red line indicates the median value, and the whiskers extend to ±2.7 standard deviations. The red asterisks indicate outliers, which reside outside of the whiskers.

Figure 5. Box and Whisker plot of root mean square (RMS) accelerometer amplitudes across all data

Figure 5. Box and Whisker plot of root mean square (RMS) accelerometer amplitudes across all data sets during rest (left side of column) and motion (right side of column). RMS values were calculated sets during rest (left side of column) and motion (right side of column). RMS values were calculated using Equation (1). The edges of the box indicate the 25th and 75th percentiles, the red line indicates using Equation (1). The edges of the box indicate the 25th and 75th percentiles, the red line indicates the median value, and the whiskers extend to ±2.7 standard deviations. The red asterisks indicate the median the whiskers to ˘2.7 standard deviations. The red asterisks indicate outliers,value, which and reside outside of theextend whiskers. outliers, which reside outside of the whiskers.

Sensors 2016, 16, 342

9 of 18

The larger the accelerometer RMS amplitude, the more intense the motion performed by the subject during testing. The accelerometer amplitude is a good indication of the level of motion introduced, but is indirectly related to the motion artifact introduced into the corrupted PPG waveform. As seen in Figure 5, about half of the data sets have a relatively small range of RMS accelerometer amplitude during motion, and median RMS accelerometer amplitudes below 2 m/s2 . Generally, these data sets showed greater differences in motion corruption between channels. However, the pressure exerted by the headband to secure the sensor and the signal amplitude of the PPG waveform, amongst other variables, can affect the quality of the recorded PPG waveforms. 3.3. PPG Motion Frequency Differences The motion frequency differences across channels were visualized by comparing Sensors 2016, 16, 342 9 of 18 the power spectral density (PSD) of the six PPG waveforms during motion with the corresponding accelerometer Sensors 2016, 16, 342Frequency Differences 9 of 18 3.3. PPG Motion signals. Figures 6 and 7 show the PSD of the PPG waveforms in all six channels, and the accelerometer The Motion motionFrequency frequencyDifferences differences across channels were visualized by comparing the power 3.3. PPG waveforms from two data sets, where channels have different PSD amplitudes of motion frequency. spectral density (PSD) of the six PPG waveforms during motion with the corresponding The motion frequency differences across channels werecorrupted visualized by comparing the power Figures 6 andaccelerometer 7 also show the MPR inthe the taken from the signals. Figuresfrequency 6 and 7 showpresent the PSD of PPG waveformsPPG in all waveform, six channels, and spectral density (PSD) of the six PPG waveforms during motion with the corresponding the accelerometer waveforms from two by dataasets, where channelsand havethe different PSD frequency amplitudes ofband present reference Masimo pulse oximeter, depicted black asterisk, motion accelerometer signals. Figures 6 and 7 show the PSD of the PPG waveforms in all six channels, and motion frequency. Figures 6 and 7 also show the MPR frequency present in the corrupted PPG in the corrupted PPG waveforms across as indicated by different the same motion frequency power the accelerometer waveforms from all twochannels data sets, where channels have PSD amplitudes of waveform, taken from the reference Masimo pulse oximeter, depicted by a black asterisk, and the motion frequency. Figures 6inand 7 also show the MPR frequency in the corrupted PPGour data sets present in themotion Acc PSD. Asband shown these figures, the majority ofpresent the motion frequency present in the corrupted PPG waveforms across all channels aspresent indicatedin by waveform, taken from the reference Masimo pulse oximeter, depicted by a black asterisk, and the the same motion frequency power present in the Acc PSD. As shown these figures, the majority of occurred in the x-direction as expected, which coincides with thein vertical up/down direction of the motion frequency band present in the corrupted PPG waveforms across all channels as indicated by the motion present in our data sets occurred in the x-direction as expected, which coincides with the subject’s bouncing the frequency exercisepower ball. present in the Acc PSD. As shown in these figures, the majority of the sameon motion vertical up/down direction of the subject’s bouncing on the exercise ball. the motion present in our data sets occurred in the x-direction as expected, which coincides with the vertical up/down direction of the subject’s bouncing on the exercise ball.

Figure 6. Power Spectral Density (PSD) plots of photoplethysmogram (PPG) waveforms from all six

Figure 6. Power Spectral Density (PSD) plots (PPG) waveforms from all six channels, and tri-axial accelerometer signals of for photoplethysmogram Data Set 9. Figure 6. Power Spectral Density (PSD) plots of photoplethysmogram (PPG) waveforms from all six channels, and tri-axial accelerometer signals for Data Set 9. channels, and tri-axial accelerometer signals for Data Set 9.

Figure 7. Power Spectral Density (PSD) plots of photoplethysmogram (PPG) waveforms from all six channels, and tri-axial accelerometer signals for Data Set 10. Figure 7. Power Spectral Density (PSD) plots of photoplethysmogram (PPG) waveforms from all six Power Spectral Density (PSD) plots of photoplethysmogram (PPG) waveforms from channels, and tri-axial accelerometer signals Data Set 10. Figure 6 shows the PR frequency and thefor motion frequency in the PSD during motion for Data

Figure 7. channels,Set and tri-axial1accelerometer signals for Setat10. 9. Channels and 3 have less amplitude in Data the PSD the motion frequency than the rest of the

Figure 6 shows the PR frequency and the motion frequency in the PSD during motion for Data channels. Set 9. Channels 1 and 3 have less amplitude in the PSD at the motion frequency than the rest of the channels.

all six

Sensors 2016, 16, 342

10 of 18

Figure 6 shows the PR frequency and the motion frequency in the PSD during motion for Data Set 9. Channels 1 and 3 have less amplitude in the PSD at the motion frequency than the rest of the channels. Sensors 2016, 16, 342 10 of 18 Figure 7 shows the PR frequency and the motion frequency in the PSD during motion for Data Set Figure 7 shows PR frequency and thein motion frequency in the PSD during Data 10. Channels 3 and 4 havethe high PSD amplitude the motion frequency, while themotion rest offor the Sensors 2016, 16, 342 10 of channels 18 Set 10. Channels 3 and 4 have high PSD amplitude in the motion frequency, while the rest of the are generally lower in PSD amplitude at the motion frequency. channels are7 generally lower in PSD amplitude at the motion frequency. shows PR frequency and the motion frequency the PSD during motion for Datamotion FigureFigure 8 shows thethe entire spectrogram from Data Set14, 14,in Channel 5. The prominent Figure 8 shows the entire spectrogram from Data Set Channel 5. The prominent Set 10. Channels 3 and 4 have high PSD amplitude in the motion frequency, while the rest motion of the frequency, as seen the Acc PSD, is 1.72 frequency, as from seen from the Acc PSD, is 1.72Hz. Hz. channels are generally lower in PSD amplitude at the motion frequency. Figure 8 shows the entire spectrogram from Data Set 14, Channel 5. The prominent motion frequency, as seen from the Acc PSD, is 1.72 Hz.

Figure 8. Spectrogram of photoplethysmogram (PPG) waveforms from Data Set 14 Channel 5. The

Figure 8. Spectrogram of photoplethysmogram (PPG) waveforms from Data Set 14 Channel 5. The dominant motion frequency appears around 1.72 Hz. Brighter coloration indicates higher power. dominant motion frequency appears around 1.72 Hz. Brighter coloration indicates higher power. Figure 8. Spectrogram of photoplethysmogram (PPG) waveforms from Data Set 14 Channel 5. The Figure 9 shows the PSD centered around the dominant motion frequency for each channel. dominant motion frequency appears around 1.72 Hz. Brighter coloration indicates higher power.

From 9these plots, is evident that around the motion power is higher in Channels 3 and 4 and From Figure shows theit PSD centered thefrequency dominant motion frequency for each channel. lower in Channels 1, 2, and 6. these plots,Figure it is evident the motion is higher Channelsfor3 each and 4channel. and lower in 9 showsthat the PSD centeredfrequency around thepower dominant motioninfrequency From plots, Channels 1,these 2, and 6. it is evident that the motion frequency power is higher in Channels 3 and 4 and lower in Channels 1, 2, and 6.

Figure 9. Spectrogram of photoplethysmogram (PPG) waveforms across all six channels from Data Set 14 centered around the dominant motion frequency at 1.72 Hz, circled in white. Brighter coloration indicates higher power. Figure 9. Spectrogram of photoplethysmogram (PPG) waveforms across all six channels from Data Figure 9. Spectrogram of photoplethysmogram (PPG) waveforms across all six channels from Data Set Set 14 centered around the dominant motion frequency at 1.72 Hz, circled in white. Brighter 14 centered around the dominant motion frequency at 1.72 Hz, circled in white. Brighter coloration coloration indicates higher power.

indicates higher power.

Sensors 2016, 16, 342

11 of 18

Sensors 2016, 16, 342

11 of 18

3.4. Multichannel Multichannel Noise Noise Level Level (MCNL) (MCNL) 3.4. Sensors 2016, 16, 342

11 of 18

Figure 10 10 shows PPG waveforms waveforms overlaid overlaid from from all all six six channels during rest rest (left) (left) and and Figure shows sample sample PPG channels during 3.4. Multichannel Noise Level (MCNL) motion (right). During rest, the majority of beats in a window are retained to form a template from motion (right). During rest, the majority of beats in a window are retained to form a template from the average average of 10 theshows “good” beats in in the thewaveforms 12 ss window, window, shown in black. black. During motion, inrest this(left) particular the of the “good” beats 12 shown in motion, this particular Figure sample PPG overlaid from all sixDuring channels duringin and window, only 7 to 10 beats were kept as “good” beats, and the template from the previous window window, 7 toDuring 10 beats were as “good” and theare template from the previous window motiononly (right). rest, the kept majority of beatsbeats, in a window retained to form a template from was used, as by tracing. average of the “good” beatsbold in the 12 s window, shown in black. During motion, in this particular wasthe used, as shown shown by the the dark dark bold tracing. window, only 7 to 10 beats were kept as “good” beats, and the template from the previous window was used, as shown by the dark bold tracing.

(a)

(b)

(a)

(b) Figure 10. Beat Beatdetection detectionand and overlay for Multichannel our Multichannel Photoplethysmogram device Figure 10. overlay for our Photoplethysmogram (MCP)(MCP) device during during (a) rest and (b) motion. Figure 10. Beat detection and overlay for our Multichannel Photoplethysmogram (MCP) device (a) rest and (b) motion. during (a) rest and (b) motion.

During motion, some channels have higher correlations with the template than other channels, During motion, some channels have higher correlations with the template than other channels, motion, some channels have higher correlations the beats template than other channels, yieldingDuring different MCNL values for each channel. Figure 11with shows overlaid in the window yielding different MCNL values for each channel. Figure 11 shows beats overlaid in the window yieldingby different for each Figure 11 shows beats overlaid window separated channelMCNL duringvalues rest (top row)channel. and motion (bottom row). During rest,initthe is evident as separated by channel during rest row) and row). During Duringrest, rest,it itisisevident evident separated channel during rest(top (topare row) andmotion motion (bottom (bottom asas expected thatby beats across all channels highly correlated with row). one another and highly correlated expected that beats across channels are highly with one oneanother anotherand andhighly highlycorrelated correlated expected that beats acrossall channels highly correlated correlated with the template shown inallFigure 11. are In contrast, during with motion, although some channels have with thethe template shown In contrast, during motion, although although somechannels channelshave have with template shownin inFigure Figure11. 11. contrast, during motion, beats that remain highly correlated withIn the template, other channels are some highly corrupted by beats thatthat remain highly correlated with the template, otherother channels are highlyhighly corrupted by motion beats remain highly correlated the template, channels corrupted by motion artifact. Therefore, the averagewith correlation coefficient and MCNLare differ between channels, artifact. Therefore, the average coefficient and MCNL between channels, allowing motion artifact. Therefore, thecorrelation average correlation coefficient anddiffer MCNL differ between channels, allowing the algorithm to select the best possible channel for PR calculations. theallowing algorithm select thetobest possible channel for PR calculations. thetoalgorithm select the best possible channel for PR calculations.

Figure Beats windowoverlaid overlaidfrom fromeach each individual individual channel Figure 11. Beats inin window overlaid from each individual channel during duringrest rest(top) (top)and and motion Figure 11.11. Beats in window channel during rest (top) andmotion motion (bottom). The template used is shown by the black trace, and the Multichannel Noise Level (MCNL) (bottom). The template template used used is is shown shown by by the the black black trace, trace, and the the Multichannel Multichannel Noise Noise Level Level (MCNL) (MCNL) (bottom). The each channel is shown thedashed dashedblack blackline. line. for for each channel is shown shown by for each channel is bybythe the dashed black line.

Sensors 2016, 16, 342 Sensors 2016, 16, 342

12 of 18 12 of 18

TheSensors overall across all six channels was averaged, and the corresponding Box-and-Whisker 2016,MCNL 16, 342 12 of 18 The overall MCNL across all six channels was averaged, and the corresponding plots of Box-and-Whisker the noise level during restnoise (left) and motion are plotted in are Figure 12.in Figure 12. plots of the level during rest (right) (left) and motion (right) plotted The overall MCNL across all six channels was averaged, and the corresponding Box-and-Whisker plots of the noise level during rest (left) and motion (right) are plotted in Figure 12.

Figure 12. Box and Whisker plots of the average Multichannel Noise Level (MCNL) across all data

Figure 12. Box and Whisker plots of the average Multichannel Noise Level (MCNL) across all data sets. The average MCNL across all channels during rest and motion are shown on the left and right Figure 12. of Box andcolumn, Whisker plots of theThe average Noise Level (MCNL) across sets. The average MCNL across all channels during andindicate motion are shown on theall leftdata and right hand side each respectively. edgesMultichannel ofrest the box the 25th and 75th percentiles, sets. The average MCNL across all channels during rest and motion are shown on the left and hand side column,therespectively. of theextend box indicate the 25th and 75th percentiles, the of redeach line indicates median value,The and edges the whiskers to ±2.7 standard deviations. Theright red hand sideindicate of eachoutliers, column,which respectively. The edges of whiskers. the box indicate the 25th and 75th percentiles, asterisks reside outside of the the red line indicates the median value, and the whiskers extend to ˘2.7 standard deviations. The red the red line indicates the median value, and the whiskers extend to ±2.7 standard deviations. The red asterisks indicate outliers, which reside outside of the whiskers. asterisks indicate outliers, which reside outside of the whiskers.

Figure 13. Multichannel noise level (MCNL) during rest and motion in Data Set 20. Accelerometer data are plotted below the MCNL to indicate where motion occurs. Figure 13. Multichannel noise level (MCNL) during rest and motion in Data Set 20. Accelerometer Figure 13. Multichannel noise level (MCNL) during rest and motion in Data Set 20. Accelerometer data are plottednoise belowlevel the MCNL to indicate where motion occurs. how clean or corrupted different Multichannel (MCNL) plots were used to illustrate

data are plotted below the MCNL to indicate where motion occurs.

channels were during motion based on their respective signal morphology. During clean segments Multichannel noise level (MCNL) plots were used to illustrate how clean or corrupted different of the PPG waveforms recorded during rest, beats across all channels showed a relatively high channels were during motion based on their respective signal morphology. During clean segments degree of correlation (C) and low MCNL The differences in the medians ranges of the Multichannel noise level (MCNL) plotsvalues. were used to illustrate how cleanand or corrupted different of the PPG waveforms recorded during rest, beats across all channels showed a relatively high MCNL, calculated by the multichannel template-matching algorithm during motion, showed high channelsdegree wereofduring motion based on their respective signal morphology. During clean segments of correlation (C) and low MCNL values. The differences in the medians and ranges of the acrossrecorded data sets, hence displaying the across variety all of motions manifested by experimental the PPGvariation waveforms during rest,template-matching beats channels a our relatively MCNL, calculated by the multichannel algorithm showed during motion, showedhigh high degree of correlation (C) and low MCNL values. The differences in the medians and ranges of the variation across data sets, hence displaying the variety of motions manifested by our experimentalMCNL,

calculated by the multichannel template-matching algorithm during motion, showed high variation across data sets, hence displaying the variety of motions manifested by our experimental protocol. The data sets with low MCNL values during motion tended to have low accelerometer RMS amplitudes. Figures 13 and 14 show the time-series of MCNL across two different data sets, indicating low values

Sensors 2016, 16, 342

13 of 18

Sensors 2016, 16, 342

13 of 18

protocol. The data sets with low MCNL values during motion tended to have low accelerometer

during rest, high values during motion, and varying values across channels during motion. In Data RMS amplitudes. Figures 13 and 14 show the time-series of MCNL across two different data sets, Set 20, the channels consistently separated termsand of noise level, whereas in Data Set 24, indicating lowwere valuesmore during rest, high values duringin motion, varying values across channels the noiseduring levelsmotion. acrossInchannels were during motion. The filtered accelerometer signal is shown Data Set 20, the closer channels were more consistently separated in terms of noise level, in Data the noise levels across channels were closer during motion. The filtered beneathwhereas the MCNL plotSet for24,motion reference. accelerometer signal is shown beneath the MCNL plot for motion reference.

Figure 14. Multichannel noise level (MCNL)during during rest for for Data Set 24. Figure 14. Multichannel noise level (MCNL) restand andmotion motion Data SetAccelerometer 24. Accelerometer are plotted the MCNL to indicatewhere where motion motion occurs. data aredata plotted belowbelow the MCNL to indicate occurs.

3.5. PR Error During Motion

3.5. PR Error During Motion

Pulse rate errors during motion were calculated in comparison to the Masimo reference sensor

Pulse rate errors during motion were calculated in comparison to the Masimo reference according to three separate parameters: performance index, accuracy, and precision. The sensor multichannel estimateparameters: (MC) corresponds to the PR measurements by switching according to three separate performance index, accuracy,taken and precision. Thebetween multichannel every 2 s usingto thethe multichannel template-matching algorithm. Databetween sets 21 and 31 showed estimatechannels (MC) corresponds PR measurements taken by switching channels every 2 s extremely low SNR, high accelerometer amplitudes during motion, high MCNL during motion, and using the multichannel template-matching algorithm. Data sets 21 and 31 showed extremely low SNR, high PR errors across all channels. Furthermore, the PPG waveforms across all channels were high accelerometer amplitudes during motion, high MCNL during motion, and high PR errors across completely corrupted by motion artifacts; therefore, these data sets were eliminated from the all channels. Furthermore, statistics calculations. the PPG waveforms across all channels were completely corrupted by motion artifacts; therefore, these data sets were eliminated from the statistics calculations. 3.5.1. PR Performance Index (PI)

3.5.1. PR Performance Index (PI) Table 1 summarizes the performance index calculated for each data set across all six channels, for 1the estimate of the the performance mean PR andindex the median PR across all data six channels, andallfor Table summarizes calculated for each set across sixthe channels, multichannel-switching PR estimates. The mean and standard error of the difference between the for the estimate of the mean PR and the median PR across all six channels, and for the multichannel, each individual channel, and the median and mean PR estimates are also shown. multichannel-switching PRcompared estimates. The mean and standard error of the difference between the Statistical significance to the multichannel-switching estimate is indicated by an asterisk multichannel, each individual channel, and the median and mean PR estimates are also shown. next to each mean PI. Statistical significance compared to the multichannel-switching estimate is indicated by an asterisk Performance index calculated for each individual channel, for the mean and median of PR next to eachTable mean1. PI. measurements, and for the multichannel-switching estimation (MC) during motion.

Table 1. Performance index for individual meanMedian and median Ch 1 calculated Ch 2 Cheach 3 Ch 4 Chchannel, 5 Ch 6for the Mean MC of PR Mean (all) 57.8% 56.2% 53.3% 54.9% 56.4% (MC) 57.9% 48.8% 55.6% 66.4% measurements, and for the multichannel-switching estimation during motion. Mean (excluding 21 and 31) Mean diff.

Mean (all) Mean (excluding 21 and 31) Mean diff. Std Err diff.

* 61.7%

* 59.9%

* 56.7%

Ch 1 9.24%

Ch 2 11.02%

Ch 3 12.35%Ch 4 10.78% Ch 59.15% Ch18.85% 6 Mean 14.15% 11.66% Median

57.8% * 61.7% 9.24% 3.05%

56.2% * 59.9% 11.02% 3.39%

53.3% * 56.7% 14.15% 4.94%

* 58.5%

* 60.1%

54.9% * 58.5% 12.35% 4.99%

* 61.7%

56.4% * 60.1% 10.78% 3.98%

* 52.0%

57.9% * 61.7% 9.15% 2.53%

* 59.2%

48.8% * 52.0% 18.85% 4.56%

70.9%

55.6% * 59.2% 11.66% 3.55%

MC 66.4% 70.9%

* indicates statistical significant compared to the multichannel-switching estimate.

We found that the performance index for the multichannel-switching estimate was on average 9.2% better than the channel with the highest average performance index and 13.6% better than the

Sensors 2016, 16, 342

14 of 18

channel with the lowest average performance index. The performance index was increased by 18.9% when compared to taking the mean of PR estimates across all channels. The multichannel-switching PI was statistically significantly better than all six individual channels, the mean PR, and the median PR. 3.5.2. PR Accuracy PR accuracy was calculated for all six channels, for the mean PR, for the median PR, and for the corresponding multichannel switching estimate. Table 2 summarizes the accuracy calculated for each data set across all six channels, the estimate of the mean PR and the median PR across all six channels, and the multichannel-switching PR estimates. Statistical significance compared to the multichannel-switching estimate is indicated by an asterisk next to each mean accuracy measurement. Table 2. Accuracy of PR for each individual channel, for the mean and median of PR measurements across channels, and for the multichannel-switching estimate (MC).

Mean (all) Mean (excluding 21 and 31) Mean diff. Std Err diff.

Ch 1

Ch 2

Ch 3

Ch 4

Ch 5

Ch 6

Mean

Median

MC

9.7 * 8.1 2.0 1.0

9.5 * 8.0 1.9 0.7

10.2 * 8.8 2.7 1.1

9.6 * 8.1 2.0 1.1

9.4 * 7.9 1.8 0.6

9.6 * 8.1 2.0 0.9

9.3 * 7.8 1.7 0.5

9.0 * 7.5 1.3 0.5

7.7 6.1

* indicates statistical significant compared to the multichannel-switching estimate.

We found that the multichannel-switching PR estimate had the highest accuracy of PR during motion, corresponding to 1.8 bpm lower than the channel with the lowest mean error and 2.7 bpm lower than the channel with the highest mean error. In addition, we found that the multichannel-switching estimate was statistically more accurate than any individual single channel estimate, and the mean and median PR across channels for all data sets recorded during motion. 3.5.3. PR Precision PR precision was calculated for all six channels, for the mean PR, for the median PR, and for the corresponding multichannel-switching estimate. Table 3 summarizes the precision calculated for each data set across all six channels, the estimate of the mean PR and the median PR across all six channels, and the multichannel-switching PR estimates. Statistical significance compared to the multichannel-switching estimate is indicated by an asterisk next to the mean precision. Table 3. Precision of PR for each individual channel, for the mean and median of PR measurements across channels, and for the multichannel-switching estimate (MC).

Mean (all) Mean (excluding 21 and 31) Mean diff. Std Err diff.

Ch 1

Ch 2

Ch 3

Ch 4

Ch 5

Ch 6

Mean

Median

MC

5.7 5.6 0.0 0.4

6.4 6.3 0.7 0.5

7.3 * 7.3 1.7 0.8

6.6 6.6 0.9 0.8

6.5 6.5 0.9 0.5

5.7 5.7 0.1 0.4

5.4 5.3 ´0.3 0.5

5.8 5.8 0.2 0.5

5.7 5.6

* indicates statistical significant compared to the multichannel-switching estimate.

The data showed that the precision of the multichannel-switching estimate was equal to the precision of the channel with the highest precision, and better than the channel with the lowest precision by 1.7 bpm. Likewise, the multichannel estimate was statistically lower than Channel 3 in precision, but not other channels. 4. Discussion Motion artifacts are the primary limiting factor in the utilization of photoplethysmography for mobile health applications. Motion artifacts are hard to quantify and filter out given the unpredictable nature of motion-induced PPG signal corruption. In this paper we obtained random, aperiodic, motion corrupted data from the forehead using a reflectance-mode, multichannel photoplethysmograph, and

Sensors 2016, 16, 342

15 of 18

introduced a multichannel-switching algorithm based on previously developed template-matching algorithms. We hypothesized that the PPG waveforms in each channel during random motion would be affected differently by motion artifacts, and that the multichannel-switching algorithm would outperform single channel estimates during motion in terms of PR error and motion tolerance. Data analysis showed that channels were different in PSD amplitude at the motion frequency. Across all data sets, depending on the severity of motion, PPG waveforms in the time domain were visually different during motion, as shown in Figure 6. Differences in amplitude at the motion frequency across channels corroborate the benefits of multichannel pulse oximetry. In particular, in the case when the motion amplitude overwhelmed a single channel, but did not affect all six channels as severely, PR measurements can still be obtained from the cleanest channel during motion. Of particular interest, when the motion frequency overlaps with the PR frequency and would be difficult to filter out if a single IR channel pulse oximeter was used, when motion amplitude did not affect all six channels to the same extent, we found that PR can still be extracted from the cleanest channel during motion with sufficient clinical accuracy. Data sets 10, 11, 16, 17, 18, 19, 20, 22, and 23 showed significant improvement from individual channels compared to the multichannel estimates in absolute PR error—one or more individual PPG channels were above the accepted tolerance in absolute PR error, while the multichannel-switching estimate was at or below tolerance in absolute PR error. The accelerometer data shows that a wide variety of motion amplitudes were introduced across all 31 data sets. Of the nine data sets where multichannel switching decreased PR error significantly, Data Sets 18, 19, 22, and 23 had a high RMS accelerometer amplitude. Hence, accelerometer amplitude is not an accurate measure of the level of motion artifact corruption present in the PPG waveform, or an indication of how the multichannel-switching approach will affect estimated PR errors during motion. The Box and Whisker plots of the MCNL showed a wide range of MCNL values for Data Sets 10, 11, 17–20, 22, and 23. These data sets showed significant improvement when the multichannel switching estimate was implemented. We found that the multichannel approach shows the most improvement when channels differ significantly in signal quality and morphology, resulting in a high variance of MCNL values during motion. Particularly, shown in Figure 11, when the MCNL is low for some channels during motion and high for other channels, the multichannel switching algorithm can choose automatically the channel with the least amount of motion corruption from which to calculate the most accurate PR values. Although the benefit in performance index during motion in term of PR error differed between data sets, when all data sets are considered, the multichannel switching estimate performed better in performance index over every individual channel by at least 9%. The benefit of multichannel-switching in PR accuracy during motion also varies across data sets, but was an improvement when all data sets were considered over every individual channel by at least 1.9 bpm during motion. For both performance index and accuracy parameters, we found that the multichannel-switching estimate was statistically significantly better than all 6 individual channels, and from both the mean PR estimates and the median PR estimates. The multichannel-switching algorithm did not outperform individual channels, or the mean or median PR across channels in precision, but was not worse in comparison to any individual channel. Therefore, the precision is the same as a standard pulse oximeter, but we believe that the benefit in performance index and accuracy achieved is enough to merit the multichannel-switching approach. The motion frequency during the majority of data sets remained relatively constant, and the majority of the motion remained in the x-direction. This led to the PPG channels with low-error PR remaining the same for the duration of motion in a number of data sets. Data sets with frequently changing motion frequency, as seen in data taken during real field motion or transportation, would implement the use of real-time channel-switching more often. And, future sensors that make use of multiple sensing sites or a number of LED-PD pairs spread out across the wrist or forehead would lead

Sensors 2016, 16, 342

16 of 18

to greater differences in channels and more diverse motion tolerance, theoretically leading to further improved PR measurements during motion by implementing our multichannel-switching algorithm. It is evident that the multichannel-switching approach presented in this paper has limitations when every channel is severely corrupted by motion artifacts. Generally, data sets with less severe motion and more significant differences in signal quality between channels will benefit more from the application of the proposed multichannel-switching algorithm during motion. Nonetheless, we think that that signal reconstruction techniques developed by our group may take advantage of the varying motion frequency content present in the six independent PPG channels to improve PR measurements during severe motion artifacts. 5. Conclusions In this paper, we explored the benefit of using a forehead-mounted multichannel photoplethysmographic sensor and investigated the performance of this wearable sensor under a variety of random, aperiodic motion. In order to investigate the feasibility of the proposed MCP sensor, we introduced an advanced multichannel-switching algorithm that selects the channel with the least amount of motion artifact to calculate PR every 2 s. We showed that for a wide variety of random motion, channels had varying amounts of signal power at the motion frequency, and that the multichannel estimate outperforms single channel estimates in terms of motion tolerance and PR errors during motion. Although the overall benefit of multichannel-switching during motion differed between data sets, the multichannel-switching estimate from the recorded PPG array outperformed each individual channel, and the mean and median of PR across channels, during motion in accuracy and performance index of PR measurements. From a practical perspective, our multichannel algorithm allows potential real-time automatic selection of the best signal fidelity channel among the six channels at each time point. Without this algorithm, a post processing stage requires a user to select which channel exhibits the minimal MA at each time point, which is a time-consuming process. These data show promise in channel switching for increased motion tolerance in PR calculations, and provide a basis for future investigation of multichannel PPG-based algorithm development. Acknowledgments: This work was supported in part by the US Army Medical Research and Materiel Command (USAMRMC) under Grant No. W81XWH-12-1-0541. Author Contributions: K.M.W. contributed in development of the algorithm and the analysis of its performance; J.R.H. participated in the data collection procedure; and K.H.C. and Y.M. contributed as research advisors. Conflicts of Interest: The authors declare no conflict of interest.

Abbreviations The following abbreviations are used in this manuscript: SpO2 PR PPG PD LED Acc MA ANC PCA ICA SSA RMS SNR MPR

Arterial Blood Oxygen Saturation Pulse rate Photoplethysmogram Photodetector Light emitting diode Accelerometer Motion Artifact Adaptive Noise Cancellation Principle Component Analysis Independent Component Analysis Singular Spectral Analysis Root Mean Square Signal-to-Noise Ratio Mean Pulse rate

Sensors 2016, 16, 342

PSD MCNL IR RD PI MC MCP

17 of 18

Power Spectral Density Multichannel Noise Level Infrared Red Performance Index Multichannel Estimate Multichannel Photoplethysmogram

References 1. 2. 3.

4.

5. 6.

7. 8. 9.

10. 11.

12.

13. 14. 15. 16.

17.

Mannheimer, P.D. The light-tissue interaction of pulse oximetry. Anesth. Analg. 2007, 105, S10–S17. [CrossRef] [PubMed] Mendelson, Y. Pulse oximetry. In Wiley Encyclopedia of Biomedical Engineering; John Wiley & Sons, Inc.: Hoboken, NJ, USA, 2006. Asada, H.H.; Hong-Hui, J.; Gibbs, P. Active noise cancellation using mems acceerometers for motion-tolerant wearable bio-sensors. In Engineering in Medicine and Biology Society; IEEE: San Francisco, CA, USA, 2004; Volume 1, pp. 2157–2160. Comtois, G.; Mendelson, Y.; Ramuka, P. A comparative evaluation of adaptive noise cancellation algorithms for minimizing motion artifacts in a forehead-mounted wearable pulse oximeter. In Engineering in Medicine and Biology Society; IEEE: Lyon, France, 2007; pp. 1528–1531. Foo, J.Y.; Wilson, S.J. A computational system to optimise noise rejection in photoplethysmography signals during motion or poor perfusion states. Med. Biol. Eng. Comput. 2006, 44, 140–145. [PubMed] Lee, B.; Han, J.; Baek, H.J.; Shin, J.H.; Park, K.S.; Yi, W.-J. Improved elimination of motion artifacts from a photoplethysmographic signal using a kalman smoother with simultaneous accelerometry. Physiol. Meas. 2010, 31, 1585–1603. [CrossRef] [PubMed] Yousefi, R.; Nourani, M.; OStadabbas, S.; Panahi, I. A motion-tolerant adaptive algorithm for wearable photoplethysmographic biosensors. IEEE J. Biomed. Health Inf. 2014, 18, 2168–2194. [CrossRef] [PubMed] Goldman, J.M.; Petterson, M.T.; Kopotic, R.J.; Barker, S.J. Masimo signal extraction pulse oximetry. J. Clin. Monit. Comput. 2000, 16, 475–483. [CrossRef] [PubMed] Ram, M.R.; Madhav, K.V.; Krishna, E.H.; Komalla, N.R.; Reddy, K.A. A novel approach for motion artifact reduction in ppg signals based on as-lms adaptive filter. IEEE Trans. Instrum. Meas. 2012, 61, 1445–1457. [CrossRef] Patterson, J.A.C.; Yang, G.-Z. Ratiometric artifact reduction in low power reflective photoplethysmography. IEEE Trans. Biomed. Circuits Syst. 2011, 5, 330–338. [CrossRef] [PubMed] Shimazaki, T.; Hara, S.; Okuhata, H.; Nakamura, H. Cancellation of motion artifact induced by exercise for ppg-based heart rate sensing. In Engineering in Medicine and Biology Society; IEEE: Chicago, IL, USA, 2014; pp. 3216–3219. Wijshoff, R.W.; Mischi, M.; Veen, J.; van der Lee, A.M.; Aarts, R.M. Reducing motion artifacts in photoplethysmograms by using relative sensor motion: Phantom study. J. Biomed. Opt. 2012, 17, 1–15. [CrossRef] [PubMed] Enriquez, R.H.; Castellanos, M.S.; Rodriguez, J.F.; Caceres, J.L.H. Analysis of the photoplethysmographic signal by means of the decomposition in principle components. Physiol. Meas. 2002, 23, N17–N29. Kim, B.S.; Yoo, S.K. Motion artifact reduction in photoplethysmography using independent component analysis. IEEE Trans. Biomed. Eng. 2006, 53, 566–568. [CrossRef] [PubMed] Krishnan, R.; Natarajan, B.; Warren, S. Two-stage approach for detection and reduction of motion artifacts in photoplethysmographic data. IEEE Trans. Biomed. Eng. 2010, 57, 1867–1876. [CrossRef] [PubMed] Peng, F.; Zhang, Z.; Gou, X.; Liu, H.; Wang, W. Motion artifact removal from photoplethysmographic signals by combining temporally constrained independent component analysis and adaptive filter. Biomed. Eng. Online 2014, 13, 1–14. [CrossRef] [PubMed] Salehizadeh, S.M.A.; Dao, D.K.; Chong, J.W.; McManus, D.; Darling, C.; Mendelson, Y.; Chon, K.H. Photoplethysmograph signal reconstruction based on a novel motion aritfact detection-reduction approach part II: Motion and noise artifact removal. Ann. Biomed. Eng. 2014, 42, 2251–2263. [CrossRef] [PubMed]

Sensors 2016, 16, 342

18.

19. 20. 21.

22. 23.

24.

25.

26. 27. 28. 29. 30.

31.

18 of 18

Chong, J.W.; Dao, D.K.; Salehizadeh, S.M.; McManus, D.D.; Darling, C.E.; Chon, K.H.; Mendelson, Y. Photoplethysmograph signal reconstruction based on a novel hybrid motion artifact detection-reduction approach. Part I: Motion and noise artifact detection. Ann. Biomed. Eng. 2014, 42, 2238–2250. [CrossRef] [PubMed] Li, K.; Warren, S. A wireless reflectance pulse oximeter with digital baseline control for unfiltered photoplethysmograms. IEEE Trans. Biomed. Circuits Syst. 2011, 6, 269–278. [CrossRef] [PubMed] Lee, J.; Jung, W.; Kang, I.; Kim, Y.; Lee, G. Design of filter to reject motion artifact of pulse oximetry. Comput. Stand. Interfaces 2004, 26, 241–249. [CrossRef] Li, K.; Warren, S.; Natarajan, B. Onboard tagging for real-time quality assesment of photoplethysmograms acquired by a wireless reflectance pulse oximeter. IEEE Trans. Biomed. Circuits Syst. 2011, 6, 54–63. [CrossRef] [PubMed] Lee, H.-W.; Lee, J.-W.; Jung, W.-G.; Lee, G.-K. The periodic moving average filter for removing motion artifacts from ppg signals. Int. J. Control Autom. Syst. 2007, 5, 701–706. Baca, A.; Biagetti, G.; Camilletti, M.; Crippa, P.; Falaschetti, L.; Orcioni, S.; Rossini, L.; Tonelli, D.; Turchetti, C. Carma: A robust motion artifact reduction algorithm for heart rate monitoring from ppg. In European Signal Processing Conference; IEEE: Nice, France, 2015. Salehizadeh, S.M.A.; Dao, D.; Bolkhovsky, J.; Cho, C.; Mendelson, Y.; Chon, K.H. A novel time-varying spectral filtering algorithm for reconstruction of motion artifact corrupted heart rate signals during intense physical activities using a wearable photoplethysmogram sensor. Sensors 2016, 16. [CrossRef] [PubMed] Vetter, R.; Rossini, L.; Ridolfi, A.; Sola, J.; Chetelat, O.; Correvon, M.; Krauss, J. Frequency domain SpO2 estimation based on multichannel photoplethysmographic measurements at the sternum. In World Congress on Medical Physics and Biomedical Engineering; Olaf Dossel, W.C.S., Ed.; Spinger: Munich, Germany, 2009; Volume 25, pp. 326–329. Wang, L.; Lo, B.P.L.; Yang, G.Z. Multichannel reflective ppg earpiece sensor with passive motion cancellation. IEEE Trans. Biomed. Circuits Syst. 2007, 1, 235–241. [CrossRef] [PubMed] Petterson, M.T.; Begnoche, V.L.; Graybeal, J.M. The effect of motion on pulse oximetry and its clinical significance. Anesth. Analg. 2007, 105, S78–S84. [CrossRef] [PubMed] Dao, D.K.; Mendelson, Y.; Chon, K.H. Multi-channel pulse oximetry for wearable physiological monitoring. In IEEE International Conference on Body Sensor Networks (BSN); IEEE: Cambridge, MA, USA, 2013; pp. 1–6. Mendelson, Y.; Ochs, B.D. Noninvasive pulse oximetry utilizing skin reflectance photoplethysmography. IEEE Trans. Biomed. Eng. 1988, 35, 798–805. [CrossRef] [PubMed] Orphanidou, C.; Bonnici, T.; Charlton, P.; Clifton, D.; Vallance, D.; Tarassenko, L. Signal-quality indices for the electrocardiogram and photoplethysmogram: Derivation and applications to wireless monitoring. IEEE J. Biomed. Health Inf. 2015, 19, 832–838. [CrossRef] [PubMed] Masimo. Masimo Radical-57: Operator’s Manual; Masimo Corporation: Irvine, CA, USA, 2012. © 2016 by the authors; licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons by Attribution (CC-BY) license (http://creativecommons.org/licenses/by/4.0/).